支持向量机的进化多核设计

Evolutionary multiple kernels design for support vector machines

为提高支持向量机分类精度,提出一种基于遗传程序设计的进化多核算法.算法中每个个体表示一个多核函数,并采用树形结构进行编码,增强了多核函数的非线性;初始种群由生长法产生,经过遗传操作后得到适合具体问题的进化多核函数.遗传程序设计的全局搜索性能使得算法设计不需要先验知识.与单核函数及其他多核函数的对比实验结果表明,进化多核有效提高了支持向量机分类性能.

To boost the classification accuracy of support-vector-machines（SVM）, we propose an algorithm with evolutionary multiple kernels（EMK）, based on the genetic programming（GP）. In this algorithm, each individual represents a multiple kernel function, and is encoded by the tree-structure for enhancing the non-linearity of the multiple kernel function. Grow method is applied to initialize the GP population, from which the EMK adapting to practical problems is obtained by genetic operations. No priori knowledge is required due to the global search of GP. Comparisons of experimental resuits of EMK with the single kernel function and other multiple kernel functions show that EMK effectively improves the classification performance of SVM.